INT8 W8A8

vLLM supports quantizing weights and activations to INT8 for memory savings and inference acceleration. This quantization method is particularly useful for reducing model size while maintaining good performance.

Please visit the HF collection of quantized INT8 checkpoints of popular LLMs ready to use with vLLM.

Note

INT8 computation is supported on NVIDIA GPUs with compute capability > 7.5 (Turing, Ampere, Ada Lovelace, Hopper).

Prerequisites

To use INT8 quantization with vLLM, you’ll need to install the llm-compressor library:

  1. $ pip install llmcompressor==0.1.0

Quantization Process

The quantization process involves four main steps:

  1. Loading the model

  2. Preparing calibration data

  3. Applying quantization

  4. Evaluating accuracy in vLLM

1. Loading the Model

Use SparseAutoModelForCausalLM, which wraps AutoModelForCausalLM, for saving and loading quantized models:

  1. from llmcompressor.transformers import SparseAutoModelForCausalLM
  2. from transformers import AutoTokenizer
  4. MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct"
  5. model = SparseAutoModelForCausalLM.from_pretrained(
  6.     MODEL_ID, device_map="auto", torch_dtype="auto",
  7. )
  8. tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

2. Preparing Calibration Data

When quantizing activations to INT8, you need sample data to estimate the activation scales. It’s best to use calibration data that closely matches your deployment data. For a general-purpose instruction-tuned model, you can use a dataset like ultrachat:

  1. from datasets import load_dataset
  3. NUM_CALIBRATION_SAMPLES = 512
  4. MAX_SEQUENCE_LENGTH = 2048
  6. # Load and preprocess the dataset
  7. ds = load_dataset("HuggingFaceH4/ultrachat_200k", split="train_sft")
  8. ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))
  10. def preprocess(example):
  11.     return {"text": tokenizer.apply_chat_template(example["messages"], tokenize=False)}
  12. ds = ds.map(preprocess)
  14. def tokenize(sample):
  15.     return tokenizer(sample["text"], padding=False, max_length=MAX_SEQUENCE_LENGTH, truncation=True, add_special_tokens=False)
  16. ds = ds.map(tokenize, remove_columns=ds.column_names)

3. Applying Quantization

Now, apply the quantization algorithms:

  1. from llmcompressor.transformers import oneshot
  2. from llmcompressor.modifiers.quantization import GPTQModifier
  3. from llmcompressor.modifiers.smoothquant import SmoothQuantModifier
  5. # Configure the quantization algorithms
  6. recipe = [
  7.     SmoothQuantModifier(smoothing_strength=0.8),
  8.     GPTQModifier(targets="Linear", scheme="W8A8", ignore=["lm_head"]),
  9. ]
  11. # Apply quantization
  12. oneshot(
  13.     model=model,
  14.     dataset=ds,
  15.     recipe=recipe,
  16.     max_seq_length=MAX_SEQUENCE_LENGTH,
  17.     num_calibration_samples=NUM_CALIBRATION_SAMPLES,
  18. )
  20. # Save the compressed model
  21. SAVE_DIR = MODEL_ID.split("/")[1] + "-W8A8-Dynamic-Per-Token"
  22. model.save_pretrained(SAVE_DIR, save_compressed=True)
  23. tokenizer.save_pretrained(SAVE_DIR)

This process creates a W8A8 model with weights and activations quantized to 8-bit integers.

4. Evaluating Accuracy

After quantization, you can load and run the model in vLLM:

  1. from vllm import LLM
  2. model = LLM("./Meta-Llama-3-8B-Instruct-W8A8-Dynamic-Per-Token")

To evaluate accuracy, you can use lm_eval:

  1. $ lm_eval --model vllm \
  2.   --model_args pretrained="./Meta-Llama-3-8B-Instruct-W8A8-Dynamic-Per-Token",add_bos_token=true \
  3.   --tasks gsm8k \
  4.   --num_fewshot 5 \
  5.   --limit 250 \
  6.   --batch_size 'auto'

Note

Quantized models can be sensitive to the presence of the bos token. Make sure to include the add_bos_token=True argument when running evaluations.

Best Practices

  • Start with 512 samples for calibration data (increase if accuracy drops)

  • Use a sequence length of 2048 as a starting point

  • Employ the chat template or instruction template that the model was trained with

  • If you’ve fine-tuned a model, consider using a sample of your training data for calibration

Troubleshooting and Support

If you encounter any issues or have feature requests, please open an issue on the vllm-project/llm-compressor GitHub repository.